burn_jit/tensor/

base.rs

use crate::element::JitElement;
use crate::kernel::{launch_unary_numeric, NumericUnaryOp, NumericUnaryOpFamily};
use crate::JitRuntime;
use burn_tensor::quantization::QTensorPrimitive;
use burn_tensor::{DType, Shape, TensorMetadata};
use cubecl::client::ComputeClient;
use cubecl::frontend::Numeric;
use cubecl::linalg::tensor::TensorHandle;
use cubecl::prelude::{TensorHandleRef, *};
use cubecl::server::Handle;
use std::marker::PhantomData;

/// The basic tensor primitive struct.
#[derive(new)]
pub struct JitTensor<R: JitRuntime> {
    /// Compute client for the [runtime](JitRuntime).
    pub client: ComputeClient<R::Server, R::Channel>,
    /// The buffer where the data are stored.
    pub handle: Handle,
    /// The shape of the tensor.
    pub shape: Shape,
    /// The device of the tensor.
    pub device: R::Device,
    /// The strides of the tensor.
    pub strides: Vec<usize>,
    pub(crate) dtype: DType,
}

impl<R: JitRuntime, E: JitElement> From<JitTensor<R>> for TensorHandle<R, E> {
    fn from(val: JitTensor<R>) -> Self {
        TensorHandle::new(val.shape.dims.to_vec(), val.strides.to_vec(), val.handle)
    }
}

impl<R> core::fmt::Debug for JitTensor<R>
where
    R: JitRuntime,
{
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.write_fmt(format_args!(
            "JitTensor {{ shape: {:?}, device: {:?}, strides: {:?}, elem: {}, runtime: {}}}",
            self.shape,
            self.device,
            self.strides,
            self.dtype.name(),
            R::name(),
        ))
    }
}

impl<R> Clone for JitTensor<R>
where
    R: JitRuntime,
{
    fn clone(&self) -> Self {
        Self {
            client: self.client.clone(),
            handle: self.handle.clone(),
            shape: self.shape.clone(),
            device: self.device.clone(),
            strides: self.strides.clone(),
            dtype: self.dtype,
        }
    }
}

impl<R: JitRuntime> TensorMetadata for JitTensor<R> {
    fn dtype(&self) -> DType {
        self.dtype
    }

    fn shape(&self) -> Shape {
        self.shape.clone()
    }
}

impl<R: JitRuntime> QTensorPrimitive for JitTensor<R> {
    fn scheme(&self) -> &burn_tensor::quantization::QuantizationScheme {
        if let DType::QFloat(scheme) = &self.dtype {
            scheme
        } else {
            panic!(
                "Quantization scheme is not valid for dtype {:?}",
                self.dtype,
            )
        }
    }
}

/// Macro to execute a kernel/operation for a given element type.
///
/// # Panics
/// Since there is no automatic type cast at this time, binary operations for different
/// floating point precision data types will panic with a data type mismatch.
#[macro_export]
macro_rules! execute_with_dtype {
    (float($dtype:expr), $element:ident, $op:expr) => {{
        match $dtype {
            burn_tensor::DType::F64 => {
                type $element = f64;
                $op
            }
            burn_tensor::DType::F32 => {
                type $element = f32;
                $op
            }
            burn_tensor::DType::F16 => {
                type $element = half::f16;
                $op
            }
            burn_tensor::DType::BF16 => {
                type $element = half::bf16;
                $op
            }
            _ => unimplemented!("Unsupported dtype"),
        }
    }};

    (float($lhs_dtype:expr, $rhs_dtype:expr), $element:ident, $op:expr) => {{
        // NOTE: might be better for floating point binary operations to return a Result instead?
        if $lhs_dtype != $rhs_dtype {
            panic!(
                "Data type mismatch (lhs: {:?}, rhs: {:?})",
                $lhs_dtype, $rhs_dtype
            );
        }
        execute_with_dtype!(float($lhs_dtype), $element, $op)
    }};
    ($dtype:expr, $element:ident, $op:expr) => {{
        match $dtype {
            burn_tensor::DType::F64 => {
                type $element = f64;
                $op
            }
            burn_tensor::DType::F32 => {
                type $element = f32;
                $op
            }
            burn_tensor::DType::F16 => {
                type $element = half::f16;
                $op
            }
            burn_tensor::DType::BF16 => {
                type $element = half::bf16;
                $op
            }
            burn_tensor::DType::U64 => {
                type $element = u64;
                $op
            }
            burn_tensor::DType::U32 => {
                type $element = u32;
                $op
            }
            burn_tensor::DType::U16 => {
                type $element = u16;
                $op
            }
            burn_tensor::DType::U8 => {
                type $element = u8;
                $op
            }
            burn_tensor::DType::I64 => {
                type $element = i64;
                $op
            }
            burn_tensor::DType::I32 => {
                type $element = i32;
                $op
            }
            burn_tensor::DType::I16 => {
                type $element = i16;
                $op
            }
            burn_tensor::DType::I8 => {
                type $element = i8;
                $op
            }
            // NOTE: bool and qfloat dtypes are actually represented as u32/u8
            // burn_tensor::DType::Bool => {
            //     type $element = u32/u8;
            //     $op
            // }
            // burn_tensor::DType::QFloat(_) => {
            //     type $element = u32;
            //     $op
            // }
            _ => unimplemented!("Unsupported dtype"),
        }
    }};
}

impl<R> JitTensor<R>
where
    R: JitRuntime,
{
    /// Create a new tensor with a contiguous memory layout.
    pub fn new_contiguous(
        client: ComputeClient<R::Server, R::Channel>,
        device: R::Device,
        shape: Shape,
        handle: Handle,
        dtype: DType,
    ) -> Self {
        let ndims = shape.num_dims();
        let mut strides = vec![0; ndims];

        let mut current = 1;
        shape
            .dims
            .iter()
            .enumerate()
            .rev()
            .for_each(|(index, val)| {
                strides[index] = current;
                current *= val;
            });

        Self {
            client,
            handle,
            shape,
            strides,
            device,
            dtype,
        }
    }

    /// Change the context of the current tensor and return the newly transferred tensor.
    pub fn to_client(
        &self,
        client: ComputeClient<R::Server, R::Channel>,
        device: R::Device,
    ) -> Self {
        let bytes = burn_common::reader::try_read_sync(
            self.client.read_one_async(self.handle.clone().binding()),
        )
        .expect("Can only change client synchronously");
        let handle = client.create(&bytes);

        Self {
            client,
            handle,
            shape: self.shape.clone(),
            strides: self.strides.clone(),
            device,
            dtype: self.dtype,
        }
    }

    /// Return the reference to a tensor handle.
    pub fn as_handle_ref(&self) -> TensorHandleRef<'_, R> {
        TensorHandleRef {
            handle: &self.handle,
            strides: &self.strides,
            shape: &self.shape.dims,
            runtime: PhantomData,
            elem_size: self.elem_size(),
        }
    }

    fn elem_size(&self) -> usize {
        if let DType::QFloat(_) = self.dtype {
            // Encoded as u32
            core::mem::size_of::<u32>()
        } else {
            self.dtype.size()
        }
    }

    /// Return the reference to a tensor argument.
    pub fn as_tensor_arg<'a, E: JitElement>(&'a self, vectorisation: u8) -> TensorArg<'a, R> {
        let handle: TensorHandleRef<'a, R> = self.as_handle_ref();

        unsafe {
            TensorArg::from_raw_parts::<E>(
                handle.handle,
                handle.strides,
                handle.shape,
                vectorisation,
            )
        }
    }

    /// Return the reference to an array argument.
    pub fn as_array_arg<E: JitElement>(&self, vectorisation: u8) -> ArrayArg<'_, R> {
        unsafe {
            ArrayArg::from_raw_parts::<E>(
                &self.handle,
                self.handle.size() as usize / core::mem::size_of::<E>(),
                vectorisation,
            )
        }
    }

    pub(crate) fn can_mut_broadcast(&self, rhs: &Self) -> bool {
        if !self.handle.can_mut() || !self.is_contiguous_buffer() {
            return false;
        }
        let ndims = self.shape.num_dims();

        for i in 0..ndims {
            let shape_lhs = self.shape.dims[i];
            let shape_rhs = rhs.shape.dims[i];

            // Output tensor will be different from the mutable tensor.
            if shape_lhs < shape_rhs {
                return false;
            }
        }

        true
    }

    /// Copy the current tensor.
    pub fn copy(&self) -> Self {
        struct Copy;

        #[cube]
        impl<N: Numeric> NumericUnaryOp<N> for Copy {
            type Options = ();

            fn execute(input: Line<N>, _options: &Self::Options) -> Line<N> {
                input
            }
        }

        impl NumericUnaryOpFamily for Copy {
            type Options<N: Numeric> = ();
            type Unary<N: Numeric> = Self;
        }

        let tensor = self.clone();

        execute_with_dtype!(
            tensor.dtype,
            E,
            launch_unary_numeric::<R, E, Copy, _>(tensor, |_| ())
        )
    }

    /// Check if the tensor is safe to mutate.
    pub fn can_mut(&self) -> bool {
        self.handle.can_mut()
    }

    /// Assert that both tensors are on the same device.
    pub fn assert_is_on_same_device(&self, other: &Self) {
        if self.device != other.device {
            panic!(
                "Both tensors should be on the same device {:?} != {:?}",
                self.device, other.device
            );
        }
    }

    /// Check if the current tensor is contiguous.
    pub fn is_contiguous(&self) -> bool {
        is_contiguous(&self.shape.dims, &self.strides)
    }

    /// Check if the current tensor has a contiguous backing buffer (no overlap and no empty memory
    /// regions within the shape).
    pub fn is_contiguous_buffer(&self) -> bool {
        self.shape.num_elements() * self.dtype.size() == self.handle.size() as usize
    }
}

pub(crate) fn is_contiguous(shape: &[usize], strides: &[usize]) -> bool {
    if shape.is_empty() {
        return true;
    }

    if shape.len() == 1 {
        return strides[0] == 1;
    }

    let mut prev_stride = 1;
    let mut current_num_elems_shape = 1;

    for (i, (stride, shape)) in strides.iter().zip(shape).rev().enumerate() {
        if i > 0 {
            if current_num_elems_shape != *stride {
                return false;
            }

            if prev_stride >= *stride {
                return false;
            }
        }

        current_num_elems_shape *= shape;
        prev_stride = *stride;
    }

    true
}

#[cfg(test)]
mod tests {
    use crate::tensor::base::is_contiguous;

    #[test]
    fn is_contiguous_basic() {
        assert!(is_contiguous(&[32, 32], &[32, 1]));
    }

    #[test]
    fn is_contiguous_permuted() {
        assert!(!is_contiguous(&[32, 32], &[1, 32]));
    }

    #[test]
    fn is_contiguous_slice() {
        assert!(!is_contiguous(&[32, 1, 64], &[32, 64, 1]));
    }

    #[test]
    fn is_contiguous_4d_positive() {
        assert!(is_contiguous(&[8, 256, 32, 32], &[262144, 1024, 32, 1]));
    }

    #[test]
    fn is_contiguous_4d_negative() {
        assert!(!is_contiguous(&[256, 8, 32, 32], &[1024, 262144, 32, 1]));
    }
}